icon-lab / TranSMS

Official Implementation of Transformers for System Matrix Super-resolution (TranSMS)

TranSMS

Official Implementation of Transformers for System Matrix Super-resolution (TranSMS)

A. Güngör, B. Askin, D. A. Soydan, E. U. Saritas, C. B. Top and T. Çukur, "TranSMS: Transformers for Super-Resolution Calibration in Magnetic Particle Imaging," in IEEE Transactions on Medical Imaging, 2022, doi: 10.1109/TMI.2022.3189693.

Demo

You can use the following links to download training, validation, test datasets.

Dataset

https://drive.google.com/drive/folders/1MvLouN4Jl1gg-Uj3PMYrq2sIw-bw_HeF?usp=sharing

Pretrained Networks:

https://drive.google.com/drive/folders/1LkElq3pQS_G_mu56lCwFgvwFysknrNXU?usp=share_link

Training

Generic training code code:

python checkTranSMSAselFFLTrain.py --useGPUno 0 --wd 0 --lr 1e-4 --scale_factor 2 --snrThreshold 5 --useNoisyProjection 1 --bs 64 --resultFolder . --n1 32 --n2 32 --trainFolder ./train --testFolder ./val

wd: weight decay, default is 0
lr: learning rate
scale_factor: 2, 4, 8, etc.
snrThreshold: SNR threshold for SM training, i.e. values below threshold are not used for training
useNoisyProjection: 0 ablates the data consistency block from TranSMS, 1 is regular TranSMS with data consistency
bs: batch size
resultFolder: path for saving model outputs
n1: SM dimension x
n2: SM dimension y
trainFolder: folder containing training SMs
testFolder: folder containing validation SMs

Code for Open MPI dataset

Code for 2x, 4x and 8x training using Open MPI dataset:

python checkTranSMSAselFFLTrain.py --useGPUno 0 --lr 5e-4 --scale_factor 2 --resultFolder . --trainFolder ./train --testFolder ./val

python checkTranSMSAselFFLTrain.py --useGPUno 0 --lr 1e-4 --scale_factor 4 --resultFolder . --trainFolder ./train --testFolder ./val

python checkTranSMSAselFFLTrain.py --useGPUno 0 --lr 5e-5 --scale_factor 8 --resultFolder . --trainFolder ./train --testFolder ./val

Inference

Code for inference using all trained networks:

python inferenceOnOpenMPI.py --useGPUno 0 --bs 256 --n1 32 --n2 32 --modelFolder ./outs/ --saveOutFolder ./results/ --testFolder ./test --interpolationMatrixPath interpolaters.mat

bs: batch size during inference
n1: SM dimension x
n2: SM dimension y
modelFolder: folder containing trained networks
saveOutFolder: path for saving "mat" file outputs
testFolder: folder containing test SMs
interpolationMatrixPath: path containing the interpolation matrix from 4x4, 8x8, 16x16 to 32x32, for fast interpolation purposes

---

Citation

You are encouraged to modify/distribute this code. However, please acknowledge this code and cite the paper appropriately.

@article{transms,
  author={Gungor, Alper and Askin, Baris and Soydan, Damla Alptekin and Saritas, Emine Ulku and Top, Can Baris; and Cukur, Tolga},
  journal={IEEE Transactions on Medical Imaging}, 
  title={TranSMS: Transformers for Super-Resolution Calibration in Magnetic Particle Imaging}, 
  year={2022},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TMI.2022.3189693}}

(c) ICON Lab 2022

Prerequisites

• Python 3.6
• CuDNN 8.2.1
• PyTorch 1.10.0

Acknowledgements

This code uses libraries from CvT, SRCNN and VDSR repositories.

For questions/comments please send an email to: [e-mail]